Three different approaches
- Rule-based chatbot: works from pre-written question-and-answer flows. Predictable but inflexible; it stalls on anything outside the script.
- General AI model (LLM): converses fluently in natural language but does not know your business; on topics it does not know, it can answer confidently and wrongly.
- RAG-based assistant: before answering, it finds the relevant information in your documents, website or database and grounds its answer in it.
What is RAG?
RAG (Retrieval-Augmented Generation) is a way of connecting a language model to your own knowledge sources. When a user asks a question, the system first retrieves the relevant passages, then the model writes its answer based on them. The assistant can then use your price list, product catalogue, processes or internal documents in their current form.
RAG’s biggest benefit is reliability: because answers are tied to sources, the assistant can show which document it relied on, and when it finds nothing it can say “I don’t know” and hand over to a person.
A good AI assistant is not one that knows everything, but one that knows what it knows and what it does not.
Which one, when?
- If a few fixed questions and redirects are enough (opening hours, address, booking link): a rule-based flow or a simple form may be more economical.
- If you need general text generation, summarising or drafting: a general model may be enough.
- If the assistant must answer accurately about your products, services, policies or documents: you need a RAG-based setup.
- If the assistant must take actions such as checking an order or creating a booking: you need tools connected to your systems (an agent setup) on top of RAG.
How does setup work?
- Scope: who the assistant helps, with which questions, and where it stops and hands over to a person.
- Knowledge sources: documents, site content, FAQs and databases are selected, cleaned and connected so they stay current.
- Rules and tone: brand voice, off-limits topics, sensitive data and answer boundaries are defined.
- Testing: accuracy, source citation and “I don’t know” behaviour are tested with real questions.
- Launch and measurement: resolved questions, handovers and unanswered questions are tracked, and the knowledge base is improved accordingly.
Data security and privacy
The information the assistant can access must be chosen carefully. Personal data should be processed only when necessary and on a lawful basis, and users should know they are talking to an AI. Which model provider is used, where data is processed and whether it is stored should be settled before setup.
Common mistakes
- Not keeping sources current: the assistant is only as accurate as the documents it draws on.
- Forgetting the human handover: leaving every question to AI can weaken the customer experience.
- Not measuring: without tracking which questions go unresolved, the assistant does not improve.
Conclusion
The right choice starts not with the technology but with the job the assistant has to do. Simple tools for simple redirects, RAG for reliable answers specific to your business, and agent setups for assistants that take action.
At SOART we build AI assistants on your own knowledge sources, in your brand’s voice and fitted to your processes. If you would like to talk about your needs, get in touch.